A kind of face information structural method for video investigationTechnical field
The present invention relates to video investigation technical field, specially a kind of face information structuring sides for video investigationMethod.
Background technology
With the fast development of science and technology, intelligent Video Surveillance Technology extensive use in police criminal detection business passes through videoVideo record pedestrian and vehicle behavior find and track the important means that suspected target has become technique of criminal investigation from video.During actually handling a case, after staff has locked suspect or suspected vehicles, needs from crime and near zone is trackedTo the course of action of suspect, however it is a very difficult job that suspect is artificially searched from the monitor video of magnanimity,Not only take longer but also easy omission target.
Invention content
The purpose of the present invention is to provide a kind of face information structural methods for video investigation, pass through face informationStructured network generates face information structural model, then obtains face information label by it, is screened for and investigates, operation letterIt is single, it is efficient and of low cost.
To achieve the above object, the embodiment of the present invention provides the following technical solutions：A kind of face for video investigation is believedStructural method is ceased, is included the following steps：
S1 obtains face from the video that needs are investigated, and marks face information；
The face information of label is carried out face characteristic extraction, and is delivered to face information structured network and is instructed by S2Practice, and generates face information structural model；
S3 passes through the face information structural model, it would be desirable to which the face information for carrying out structuring is sent to face letterIt ceases in structured network, to obtain the corresponding face information label of face information；
S4 carries out face screening and investigation according to obtained face information label.
Further, whether in the S1 steps, the face information of label includes at least gender, age, wear glasses and skinColor.
Further, the S2 steps are specially：
The face characteristic of extraction is delivered in face information structured network, is carried out by concatenated convolutional neural network specialSign extraction, and carry out target type classification；
The face information data marked are divided into three subsets, and in proportion 4:4:2 are combined at random, mark respectivelyFor the first face set, the second face set and third face set；
Meanwhile respectively by the first face set, the second face set and the third face set send toFace information structured network is trained, and generates face information structural model.
Further, it carries out target type classification and uses cross entropy loss function,
A certain face is denoted as xi, corresponding loss function is：
It is the output probability of face kth kind information, φ represents the network parameter of face structuring,Indicate k kinds letterThe tag along sort of breath.
Further, the first face set, the second face set and the third face set have eight respectively100000,800,000 and 400,000 face informations.
Further, the face information structured network is by two 37 layers of convolutional neural networks and five layers of convolutionNeural network cascades.
Further, S2 steps described in repetitive operation can be obtained by roughly to fine face information structural model.
Compared with prior art, the beneficial effects of the invention are as follows：
1, face information structural model is generated using face information structured network, face information mark is then obtained by itLabel directly carry out the face sample labeling in video monitoring to face information, not only easy to operate, efficient, of low cost, andAnd it is more preferable to the environmental suitability of actual video investigation, robustness is stronger, is more practically applicable under battle conditions.
2, it using the structuring of face information, can solve to search suspect from the monitor video of magnanimity in the prior artAbnormal difficult defect.
3, by concatenated convolutional neural network method, the information of face picture has been carried out by roughly to fine screening,More accurate human face structure is obtained as a result, high degree helps criminal detective to screen suspicious object, shortening investigation time.
Description of the drawings
Fig. 1 is a kind of step flow of face information structural method for video investigation provided in an embodiment of the present inventionFigure；
Fig. 2 is a kind of structuring grade of face information structural method for video investigation provided in an embodiment of the present inventionJoin the Organization Chart of neural network；
Fig. 3 is a kind of face information of face information structural method for video investigation provided in an embodiment of the present inventionMark schematic diagram；
Fig. 4 is a kind of cascaded neural of face information structural method for video investigation provided in an embodiment of the present inventionThe illustraton of model of network；
Specific implementation mode
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, completeSite preparation describes, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based onEmbodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts all otherEmbodiment shall fall within the protection scope of the present invention.
Referring to Fig. 1, the embodiment of the present invention provides a kind of face information structural method for video investigation, including such asLower step：S1 obtains face from the video that needs are investigated, and marks face information；S2 carries out the face information of labelFace characteristic extracts, and is delivered to face information structured network and is trained, and generates face information structural model；S3,Pass through the face information structural model, it would be desirable to which the face information for carrying out structuring is sent to face information structured networkIn, to obtain the corresponding face information label of face information；S4 carries out face screening according to obtained face information label and detectsIt looks into.Face information structural model is generated using face information structured network, face information label is then obtained by it, directlyFace sample labeling in video monitoring is carried out to face information, it is not only easy to operate, efficient, of low cost, but also to realityThe environmental suitability of border video investigation is more preferable, and robustness is stronger, is more practically applicable under battle conditions.And the structuring of face information is used,It can solve to search the extremely difficult defect of suspect from the monitor video of magnanimity in the prior art.
As the prioritization scheme of inventive embodiments, in S1 steps, the face information of label include at least gender, the age,Whether wear glasses and the colour of skin.Other feature information, such as height can also be added according to actual conditions.Label for labelling example is such asShown in Fig. 3, the present invention acquires and is labelled with 2,000,000 human face data collection.
As the prioritization scheme of inventive embodiments, as shown in Fig. 2, S2 steps are specially：The face characteristic of extraction is conveyedInto face information structured network, feature extraction is carried out by concatenated convolutional neural network, and carry out target type classification；It willThe face information data marked are divided into three subsets, and in proportion 4:4:2 are combined at random, are respectively labeled as the firstFace set, the second face set and third face set；Meanwhile respectively by the first face set, second faceSet and the third face set are sent to face information structured network and are trained, and generate face information structuring mouldType.In the present embodiment, the first face set, the second face set and third face set are corresponding with 800,000, eight respectively100000 and 400,000 face informations.
As the prioritization scheme of inventive embodiments, carries out target type classification and use cross entropy loss function, by a certain peopleFace is denoted as xi, and corresponding loss function is：It is face kth kind letterThe output probability of breath, φ represent the network parameter of face structuring,Indicate the tag along sort of k kind information.It is damaged using cross entropyFunction is lost to calculate human face structure network parameter, can accurately and quickly be obtained a result.
As the prioritization scheme of inventive embodiments, referring to Fig. 4, face information structured network is rolled up by two 37 layersProduct neural network and five layers of convolutional neural networks cascade.The cascade network mainly has convolutional layer, pond layer and connects entirelyConnect the connections such as layer composition.
As the prioritization scheme of inventive embodiments, S2 steps described in repetitive operation can be obtained by roughly to fine faceMessage structure model.In S3 steps, step 2 can obtain a face information structural model from thick to thin, pass throughTrained face information structural model is sent to the face for needing to carry out structuring in prediction network, can obtain peopleFace message structureization corresponding classification results need a large amount of human face datas analyzed when video investigation, pass through cascade network modelAfter obtaining face information classification, the face information structuring, it can be achieved that 2,000,000 is stored to classification information.
As the prioritization scheme of inventive embodiments, in S4 steps, according to step 3, can obtain under big data environmentHuman face structure information can carry out quickly excluding and screening according to these label informations, convenient for target in video investigationQuickly investigation.
It although an embodiment of the present invention has been shown and described, for the ordinary skill in the art, can be withUnderstanding without departing from the principles and spirit of the present invention can carry out these embodiments a variety of variations, modification, replaceAnd modification, the scope of the present invention is defined by the appended.